OPEN FIELD DATABASE / VERSION 1

Agent failures worth routing to a human.

A public, practical index of failure patterns observed in automated research and tool workflows. Use it to decide when an agent should stop, document uncertainty, and request human judgment.

1. Source collapse

   Signal: The agent cites a page that repeats a claim but does not prove it.

   Safer route: Evidence Check

2. Tool-state mismatch

   Signal: The agent assumes a tool action succeeded when the external system never confirmed it.

   Safer route: Exception Triage

3. Stale fact

   Signal: A once-correct fact changed, but the agent relies on an old source or memory.

   Safer route: Evidence Check

4. Permission boundary

   Signal: The next action would exceed the operator's approved tools, data, or spending authority.

   Safer route: Human authorization

5. Ambiguous identity

   Signal: Two companies, people, records, or accounts appear to be the same but are not verified.

   Safer route: Evidence Check

6. Looping recovery

   Signal: The agent repeats the same failed step without changing the hypothesis.

   Safer route: Exception Triage

7. Partial success

   Signal: A workflow completes one system step but silently fails downstream.

   Safer route: Exception Triage

8. Unsupported confidence

   Signal: The agent gives a definite recommendation even though critical evidence is missing.

   Safer route: Evidence Check

9. Sensitive-data overreach

   Signal: The task begins to require credentials or unnecessary personal information.

   Safer route: Stop and escalate

10. Human-impact decision

    Signal: The output could materially affect hiring, access, money, or another person's rights.

    Safer route: Accountable human review